Re: [PATCH v6 bpf-next 0/5] bpf_prog_pack followup

[Song Liu <songliubraving@fb.com>]

> On Jul 8, 2022, at 8:58 AM, Luis Chamberlain <mcgrof@kernel.org> wrote:
> 
> On Fri, Jul 08, 2022 at 01:36:25AM +0000, Song Liu wrote:
>> 
>> 
>>> On Jul 7, 2022, at 5:53 PM, Luis Chamberlain <mcgrof@kernel.org> wrote:
>>> 
>>> On Thu, Jul 07, 2022 at 11:52:58PM +0000, Song Liu wrote:
>>>>> On Jul 7, 2022, at 3:59 PM, Luis Chamberlain <mcgrof@kernel.org> wrote:
>>>>> 
>>>>> On Thu, Jul 07, 2022 at 03:35:41PM -0700, Song Liu wrote:
>>>>>> This set is the second half of v4 [1].
>>>>>> 
>>>>>> Changes v5 => v6:
>>>>>> 1. Rebase and extend CC list.
>>>>> 
>>>>> Why post a new iteration so soon without completing the discussion we
>>>>> had? It seems like we were at least going somewhere. If it's just
>>>>> to include mm as I requested, sure, that's fine, but this does not
>>>>> provide context as to what we last were talking about.
>>>> 
>>>> Sorry for sending v6 too soon. The primary reason was to extend the CC
>>>> list and add it back to patchwork (v5 somehow got archived). 
>>>> 
>>>> Also, I think vmalloc_exec_ work would be a separate project, while this 
>>>> set is the followup work of bpf_prog_pack. Does this make sense? 
>>>> 
>>>> Btw, vmalloc_exec_ work could be a good topic for LPC. It will be much
>>>> more efficient to discuss this in person. 
>>> 
>>> What we need is input from mm / arch folks. What is not done here is
>>> what that stuff we're talking about is and so mm folks can't guess. My
>>> preference is to address that.
>>> 
>>> I don't think in person discussion is needed if the only folks
>>> discussing this topic so far is just you and me.
>> 
>> How about we start a thread with mm / arch folks for the vmalloc_exec_*
>> topic? I will summarize previous discussions and include pointers to 
>> these discussions. If necessary, we can continue the discussion at LPC.
> 
> This sounds like a nice thread to use as this is why we are talking
> about that topic.
> 
>> OTOH, I guess the outcome of that discussion should not change this set? 
> 
> If the above is done right then actually I think it would show similar
> considerations for a respective free for module_alloc_huge().
> 
>> If we have concern about module_alloc_huge(), maybe we can have bpf code 
>> call vmalloc directly (until we have vmalloc_exec_)? 
> 
> You'd need to then still open code in a similar way the same things
> which we are trying to reach consensus on.

>> What do you think about this plan?
> 
> I think we should strive to not be lazy and sloppy, and prevent growth
> of sloppy code. So long as we do that I think this is all reasoanble.

Let me try to understand your concerns here. Say if we want module code
to be a temporary home for module_alloc_huge before we move it to mm
code. Would you think it is ready to ship if we:

1) Rename module_alloc_huge as module_alloc_text_huge();
2) Add module_free_text_huge();
3) Move set_memory_* and fill_ill_insn logic into module_alloc_text_huge()
  and module_free_text_huge(). 

Are these on the right direction? Did I miss anything important?

Thanks,
Song